Compare · 10 models
Protein design models you can run online
Computational models that generate novel protein backbones, fit sequences to structural scaffolds, and design targeted binders for structural biology and protein engineering applications.
What protein design models do
Protein design spans the generation of de novo structural backbones, solving the inverse folding problem to identify amino acid sequences that adopt specific conformations, and engineering molecules to bind target surfaces. Rather than relying solely on naturally occurring sequences or random mutagenesis, these computational approaches model the biophysical and statistical relationships between sequence, structure, and functional interfaces.
Modern methods address these tasks through distinct paradigms. Denoising diffusion architectures generate backbone coordinates from noise or simultaneously produce sequence and all-atom structures. When backbone scaffolds are pre-defined, inverse folding models predict compatible residues while accommodating structural constraints. Specialised pipelines further incorporate folding backpropagation or active learning to design target binders and guide iterative laboratory screening.
Compared
- ProGen2Salesforce ResearchGenerates protein and antibody sequences conditionally using autoregressive language modelling.Commercial OKReady
- ProteinMPNNBaker LabGenerates amino acid sequences predicted to fold into specified target protein backbone structures.Commercial OKReady
- RFDiffusionBaker Lab / Institute for Protein DesignGenerates de novo protein backbones and targeted binders using SE(3) denoising diffusion.Non-commercialReady
- AntiFoldHummer et al., DTUAn antibody inverse folding model that designs CDR sequences tailored to antibody backbone frameworks.Commercial OKOn demand
- ESM-IF1Meta FAIRGenerates protein sequences for specified backbone structures using ESM2 language model representations.Commercial OKOn demand
- SaProtSu et al., Westlake UniversityA structure-aware protein language model predicting variant effects and generating embeddings using 3Di tokens.Commercial OKOn demand
- BindCraftOvchinnikov Lab / Baker LabAn automated pipeline designing target-specific protein binders via AlphaFold2 co-folding hallucination and MPNN.Commercial OKNot yet
- DPLM-2Nanjing University / ByteDanceGenerates protein sequences and structures simultaneously using a unified diffusion process.Commercial OKNot yet
- EVOLVEproHarvard / Broad InstituteActive learning model for directed evolution that proposes mutant sequences from small experimental batches.Commercial OKNot yet
- PeptiVerseChatterjeeLabA suite of machine learning models predicting biophysical and therapeutic properties of natural and modified peptides.See its pageNot yet
How to choose
For de novo backbone generation, RFDiffusion creates structural coordinates using SE(3) denoising diffusion, supporting unconditioned generation as well as conditioning on target surfaces, functional motifs, or symmetries. Because RFDiffusion outputs backbones with poly-glycine sequences in designed regions, it requires downstream sequence design with an inverse folding tool such as ProteinMPNN. In contrast, DPLM-2 unifies sequence and structure generation into a single diffusion process to prevent mismatches between fold and residue identity, outputting all-atom coordinates and sequences concurrently, though its conditioning options are restricted to specified lengths and motif sequences.
When designing sequences for pre-existing coordinates, ProteinMPNN and ESM-IF1 provide complementary inverse folding approaches. ProteinMPNN operates on full backbone atoms or alpha-carbon coordinates alone, scaling up to 200,000 residues with extensive sequence controls including residue fixing, position tying, and amino acid biasing. ESM-IF1 uses ESM2 representations rather than graph neural networks, scoring candidate sequences with perplexity and recovery metrics; it requires at least 16 GB of GPU memory and can be run alongside ProteinMPNN to establish consensus designs. For antibody-specific tasks, AntiFold is trained on antibody structures to redesign CDR loops and backbones while preserving canonical framework geometry, requiring at least 8 GB of GPU memory.
For targeted binder design, BindCraft provides an automated pipeline using AlphaFold2 co-folding hallucination, MPNN sequence redesign, and PyRosetta relaxation. It targets specific hotspots, patches, or unconstrained sites with high experimental efficiency, though it demands 32 GB to 40 GB of VRAM and exhibits reduced performance on hydrophilic target interfaces.
For downstream optimisation and variant scoring, SaProt combines primary amino acid residues with Foldseek 3Di structural tokens to predict variant effects, using pLDDT masking for regions scoring below 70 in AlphaFold 2 structures. EVOLVEpro guides directed evolution campaigns using an active learning framework built upon ESM2, proposing batches of diverse variants from empirical assay measurements with as few as 24 data points per round. For peptide engineering, PeptiVerse evaluates therapeutic and biophysical properties across both standard amino acid sequences and SMILES representations, incorporating conformal prediction intervals and predictive entropy for uncertainty quantification.
Questions
Can I run these models without a GPU or installing anything?
6 of the 10 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
How do RFDiffusion and ProteinMPNN work together?
RFDiffusion generates three-dimensional backbone coordinates from random noise but outputs designed regions as poly-glycine backbones. ProteinMPNN is subsequently applied to those generated coordinates to design compatible amino acid sequences predicted to fold into the target conformation.
Which models are specialised for antibody sequence design?
AntiFold is trained specifically on antibody structural data to redesign CDR loops and framework regions given an input coordinate file.
What is the difference between ProteinMPNN and ESM-IF1?
ProteinMPNN is a graph-based model that supports full backbone or alpha-carbon-only inputs, scales up to 200,000 residues, and permits granular control such as residue tying, fixing, and compositional biasing. ESM-IF1 uses ESM2 language model representations rather than graph networks, scoring designs by perplexity and sequence recovery while requiring at least 16 GB of GPU memory.
What are the GPU memory requirements across these design tools?
Hardware demands vary across the models: AntiFold requires at least 8 GB of GPU memory; ESM-IF1, SaProt, and EVOLVEpro require a minimum of 16 GB; ProGen2 and DPLM-2 require 24 GB; and BindCraft requires between 32 GB and 40 GB of VRAM for larger target and binder complexes.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.